DataPacts

Seller track · Lesson 1 of 9 · 10:11

Is your company's data worth selling?

What you'll learn

  • What a yes really commits you to
  • Six questions before you talk to any buyer
  • How to read the numbers in buyer ads
  • When no is the right answer

Transcript

Sell your company's data to an AI company, and you can't take it back. One data program makes its license irrevocable and perpetual. A data labeling company says it plainly: a model trained on your data can't unlearn it. So before anyone talks price, ask a simpler question. Should you sell at all?

First, what a yes actually commits you to.

Saying yes usually means granting a license to train. You keep your data. But that license often outlasts the whole commercial relationship, so it's worth knowing the terms before you get excited about the price.

Take one data program's published master terms, updated in November 2026. The grant is worldwide, royalty-free and nonexclusive. It's also irrevocable and perpetual. Irrevocable means you can't call it back. Perpetual means it has no end date.

What if you delete your raw data afterwards? Under section 4.4 of those terms, the buyer deletes within a hundred and twenty days after delivery, and has no duty to retrain its models. A data labeling company says it on its own FAQ: models already trained on the data cannot be unlearned.

Exclusivity is another thing you can give away. One data labeling company states its licenses are exclusive by default. A talent platform says it's the sole seller for the first ninety days. Terms can even differ inside one buyer: its agreement runs until terminated, while its FAQ says the license is perpetual.

Now the six questions.

A clear no to any of the first three ends the process before anyone mentions price. Do you hold the rights? Can the personal data come out, or be lawfully used? Can you accept a grant you can't take back? The other three come after.

Question one, rights. You can license only what you have the right to license. One chat tool's API terms, for example, bar third-party apps from using API data to train a large language model. And buyers ask you to confirm it. One data program's terms have you warrant, in section 5.1, that you haven't granted anyone the same rights on similar terms.

Question two, personal data. Customer and staff records are the most common reason a company isn't ready yet, because that data has to be removed, or covered by a lawful basis, before any license.

And removing names isn't enough. Researchers reported in Nature Communications that ninety-nine point nine eight percent of Americans would be correctly re-identified in any dataset using fifteen demographic attributes.

Get it wrong in Europe and GDPR Article 83 sets fines up to twenty million euros, or four percent of worldwide annual turnover, whichever is higher. Which rules reach your records is a question for a lawyer.

Question three is the grant you can't take back, which you've just seen. Question four is program thresholds. Programs publish entry floors. One states thirty or more full-time employees. Another states twenty or more, plus three or more years in operation. Overall the floors run from twenty to fifty employees. And meeting a floor gets you through the door, nothing more.

Read the page that governs your route, because a program can disagree with itself. One program's home page says thirty employees. Its referral page says fifty. And its agreement says fifty office-based.

Question five, would the income matter, and would it last? One stock-photo company's data, distribution and services revenue was fifteen point eight million dollars in the third quarter of 2026, against sixty-five point seven million a year earlier. An academic publisher's AI license revenue fell from twenty-eight point nine million to thirteen point seven.

Question six, have your staff and customers been told, where the rules require it? Staff rules apply in Germany under BetrVG section 87, in the Netherlands under WOR article 27, and in France under the Code du travail. In California, the CCPA exemption for employee data has been inoperative since January first, 2023. Take that one to your lawyer.

There's a practical test too. Can you even export the data on your current plan? On your chat tool's free and pro plans you can export public channels. Private channels and direct messages need a business or enterprise plan. Your drive provider's export bucket auto-deletes sixty days after the start. And your email suite's data is deleted no later than a hundred and eighty days after cancellation.

Think about the paperwork too. One program's terms require consent and release records kept for seven years after the agreement ends, and security incidents reported within two business days. And someone has to sign. Can one named person approve the scope? Can you keep records afterwards?

Put it together as a twelve question readiness check. Every no on it is a task to finish before you contact a buyer. And several nos can simply mean you don't sell.

Now the question everybody asks first. What's it worth?

The honest answer starts here. No public rate card exists. Company data is worth what an AI buyer will pay for the right to train on it, usually under a license. So every number you'll see comes with a named party behind it.

Buyer ranges are that buyer's own claims. One talent platform states listings of ten thousand to one and a half million dollars per company. One program states tiers from a hundred thousand upward. Another platform's headline says earn a hundred thousand to four million plus, while the calculator on the same page says your data could be worth up to a hundred thousand.

One data company's estimator gave four different ranges for the same inputs, in one day, four page loads. One ran from about fifty-two thousand to a hundred and twenty-two thousand dollars. Another ran from about thirty-nine thousand to ninety thousand. A calculator is one buyer's illustration. It isn't a valuation.

So how much have sellers actually been paid? The one named founder reports roughly twenty-five thousand dollars net for a codebase, in a technology magazine. Everything else is unconfirmed: no seller-confirmed amount received under any program is public.

So read any figure by who states it. An advertised range is a sales message. A claimed total is the payer's own count. A contract value or a bid is a promise on the record. Only a confirmed amount shows money actually received.

Filed deals are real, but narrow. One social platform's filing discloses two hundred and three million dollars in data licenses with unnamed counterparties. A health data company states two hundred million over three years, fifty million of it upfront. Each deal prices one data type. Clinical records and book titles set no rate for an archive of tickets or email.

The closest thing to a price for workplace records is a bankrupt US airline. Its emails, chat messages and IT tickets went up for auction, and a big search company made the successful bid: ten million dollars, for de-identified data only. The court still had to approve the sale when this lesson was made.

Six drivers move the value. Rights. Customer consent. Personal data. Exclusivity. Uniqueness. And volume, history and freshness.

Uniqueness matters: one data-cleaning company says about thirty-seven percent of the raw code offered through its marketplace is removed in cleaning. And rights, consent and personal data can each take the value to zero.

So the only number worth trusting is written offers on comparable terms from more than one buyer, built on an inventory and a rights check. And compare the terms, not just the price. One talent platform states payment typically two to four weeks after signing. Another program, about three months.

Last chapter. When the answer is no.

Not selling is a valid outcome, and it's the only choice that leaves every other path open. You can keep the data and use it internally. License a narrower, nonexclusive slice. Take revenue share, which pays only when the data sells. Or just wait, because no grant has been made.

Some paths can't be undone. Open release and deletion are the hardest to reverse. Once data is public, a buyer can usually get it without a license, and some programs reject it outright. And deleting your own copy does nothing to a model already trained on it.

So it's no, at least for now, when contracts or confidentiality clauses bar the use, when a platform's terms bar it, or when personal data can't be de-identified.

It's also a no when you're offered an exclusive, perpetual grant for a low fixed fee, or when an unsolicited approach sets a short deadline. But a no now can become a yes after the rights and consent work.

So, in one line. Hold the rights, clean the personal data, be ready for a grant you can't recall, and compare written offers before you believe any number. Saying no is a perfectly good answer too.

Next up is lesson A2, What buyers actually want. The full course is free at datapacts dot com slash academy.

Sources on DataPacts